AI doesn’t have to be capable of half of what’s been promised to make a significant impact on risk technology. You don’t need 1000 machine learning specialists; you don’t need half your staff researching neural nets. From cleaning up data to boosting process efficiency, what AI can actually do today can earn its keep.¹

In this blog, we’ll explore three ways existing AI technology can be utilized by financial institutions (FIs) to improve operations in risk management and anti-money laundering. Missed our previous posts on AI for risk? Start at the beginning of this series, then come back.

Bring KYC into the 21st Century

Modern KYC standards have FIs choosing between fires and frying pans. Compliance means long waits, costly due diligence, and the loss of potential customers. Skimping saves those headaches…in exchange for penalties. And this paradox has become so evident that governments are taking note. As Chartis Research recently reported, “In September 2016, for example, the Hong Kong Monetary Authority issued a company circular, warning against the potential negative effects of overly stringent onboarding and CDD processes.”²

While regulatory easing wouldn’t hurt, the path forward is better technology.

Natural language processing (NLP), the field of AI dedicated to analyzing human communication, is one such technology. KYC systems that combine NLP with ML can rapidly identify people, companies, organizations, and map relationships that connect them. From customer legal documents and proprietary knowledge bases, NLP/ML-powered applications can scour unstructured and structured data sources (almost regardless of language) and create summarized reports, generate detailed relationship visualizations, and offer confidence-rated resolution alternatives. An entity resolution process built in this manner would grow more accurate and efficient over time, cut wait times and compliance costs, and reduce the likelihood of penalties.

Given that information is doubling bi-annually and all but 10% of it is unstructured, the vast majority of due diligence will need to be done in a surging swamp of language.³ The globalization of business adds yet another layer of difficulty as FIs will have to become increasingly competent handling names of relatively unfamiliar origin. In other words, what is already difficult for current methods is going to approach impossible.

For these reasons and many more, AI’s multilingual, adaptive, relationship-discovering qualities make it a perfect match for FIs KYC challenges.

Improve Fraud Detection

If AI excels at anything, it’s pattern recognition. One of the strengths of unsupervised machine learning is its ability to identify patterns analysts would have never found. While the ability to spot the anomalous is universally useful, it is particularly applicable to fraud detection—where pattern recognition is the name of the game. Yet, despite being seemingly tailor-made, the use of ML is far from standard.

Most fraud detection relies on rule-based systems to flag suspicious activity. In this approach, a fraud specialist compiles and codifies a set of conditions that, if met by a particular transaction, would cause the system to alert an investigator. Transactional dollar limits are a classic example of programmatic heuristic; under such a rule, the system would note and record all transfers over a given amount alert a fraud investigator for follow-up. While some small fraction of these alerts may lead to real criminal activity, the vast majority are false positives—an unfortunate reality made more costly to FIs by the numerous hours of investigative work required to review each alert.

These rules-based systems aren’t a mystery to criminals, who are often able to camouflage their activity according to the specific transactional patterns that trigger investigation. A common response to the flood of false positives and criminal cleverness has been to load the system with even more rules, but the resulting software can—and often does—collapse under the weight of its own programming.

As ML expert Pradeep Banavara notes in “Applying Machine Learning and AI In Fraud Detection: A Primer:”

“What we need is a speedier mechanism to detect patterns and form rules without the added complexity of software mutation and changes. We need systems that can start from a basic rule and enable us to update these rules based on patterns detected from data, at a rate not possible by humans. Today, fortunately we have machine learning algorithms that excel at these tasks. These algorithms can analyze petabytes of data in a matter of hours and detect patterns with very simple computations.”

This approach to detection is not only more elegant and adaptive than rule-based systems but also capable of finding patterns of fraudulent activity undetectable by humans. It would even be possible to train such systems against simulated criminal programs designed to generate novel evasion strategies, allowing for the identification of patterns before they’ve been encountered in the wild.

Fraud is a complex crime, and FIs need every technological resource they can find to fight it. AI represents precisely the kind of resource they need. In fact, fraud detection is probably one of the most obvious applications of AI for risk.⁴

Make AML Cheaper

As was the case with fraud, most existing anti-money laundering (AML) technology is the dominion of rule-based systems. And current systems are likewise floundering. “Hand-coded rules are hopeless in the face of the torrent of data experienced by FIs” writes Ayasdi’s Jonathan Symonds. “From the volume and velocity of the financial transaction data, to its inherent complexity and unlabeled nature – banks cannot effectively distinguish between signal and noise.”⁵ As machine systems will always perform the first pass on transactional data for investigators, ML’s outperformance of rules-based approaches makes this evolution all but inevitable.

But the pitch here isn’t just performance. It’s about saving money. While FIs may use similar techniques to fight fraud and money laundering, the two efforts are rarely viewed the same way. Fraud has a direct impact on firm revenues, and this allows effective fraud departments to make a clear case for their value. Since the benefits of effective AML are society-wide, even the best departments are still cost-centers.⁶

And the costs are astronomical. According to a recent study by LexisNexis,

“the cost of AML compliance among European financial institutions in five key markets including France, Germany, Italy, Switzerland, and The Netherlands….was estimated at a staggering €70.1 billion ($83.5 billion) annually. AML compliance officers reported that these costs have risen 21 percent over the past two years, with another 17 percent increase expected in 2017.”⁷

How can FIs fight back? They can start by integrating systems that don’t flood their departments with false positives. PwC reported that 90-95% of the millions of alerts received every year by large FIs are false positives. To handle this influx, FIs spend millions on small armies of specialists to rifle through and review each alert, a process that can take hours and hit $30 per case.⁸ At such levels of inefficiency, even a minor reduction in the rate of false positives translates to millions in savings.

But the application of AI to AML means much more than a minor reduction in costs. From segmentation to transaction monitoring, machine learning techniques have logical applications throughout the AML pipeline, the integration of which would dramatically cut the number of false positives that currently plague investigators. According to the recent Risk100 report by Chartis, machine learning applications could slash the rate of spurious alerts by half. Along with improved law-enforcement, accuracy improvements of that magnitude would save FIs tens of millions.

In addition to improved detection, AI can also soften AMLs impact on the bottom line by automating compliance processes. The financial world still feels the tremors of the financial crisis, and nowhere are these effects felt more than in regulation. With MiFID II and IFRS9 live this year, sweeping reforms continue to be the norm, and NLP is a natural fit for the ever-growing mass of text that is financial regulation. Using this approach, applications can be trained on the terms and topics relevant to a given firm’s AML activities, analyze reams of regulatory updates, and report what matters to compliance officers.⁹ And this is just one example of how AI techniques could help lower the costs of compliance.

From fighting false positives to interpreting regulations, AI can help bring AML’s hefty contribution to the cost of doing business down. While implementation hurdles are inevitable, the efficiency gains FIs stand to make off of the AI integration are simply too great pass up.

Learn more

Clearly there are valuable uses for AI in the risk industry. What are the limits however, what can’t AI do? Check back next week to find out, or download the complete Honest Guide to AI for Risk now.

---

^{1 Consider Solutions, “World Class Finance: Things to Consider for 2018, There’s Something for Everyone”
2 Chartis Research, “RiskTech100® 2018”
3 Vernon Turner, “The Digital Universe of Opportunities;” Deloitte Advisory, “Why artificial intelligence is a game changer for risk management”
4 Ladd Muzzy, “How Artificial Intelligence can Influence Governance, Risk and Compliance”
5 Jonathan Symonds, “Intelligent Segmentation as the Attack Point for AML”
6 Sal Jadavji, “Fraud and Money Laundering: What’s the connection?”
7 Vamsi Koduru, “A more intelligent way to fight financial crimes”
8 FIs often receive millions of such alerts.
9 Mike MacDonagh, “How AI has the potential to transform Regulatory Compliance”}